docs(api): deepseek-v4-flash on /responses, qwen3.8-flash 1M context, two new 400s

[sre-helmcode]

What

1. /responses supports deepseek-v4-flash (description + model enum; new stream property). Streaming now documented per model: deepseek-v4-flash streams incrementally with the full event sequence; qwen3.6 and gemma4 hold the answer and send it in one burst at the end. The stale "single terminal event" sentence is gone from the overview, the /responses description and the Codex guide (EN/ES).
2. qwen3.8-flash context is 1,048,576, not 262,144. Model card, choose-a-model, cline, opencode (limit.context 1048576, output stays 32768), pi (contextWindow), vscode (maxInputTokens 1015808 = window minus output), API catalog, home table (modelos.json), modelCatalog.ts. EN and ES. Other models untouched.
3. Two new 400s:
   • Tool parameters, when present, must have root "type": "object" (any model). Tool schema pins it (required: [type], type enum ["object"]) and says so in prose.
   • Context overflow: 400, not retried automatically (errors table, shared BadRequest, max_tokens). The exact Context length exceeded for model '...' text is promised only on qwen3.6/gemma4 (measured); on deepseek-v4-flash and other external models it is "usually" that text, since a floor-induced overflow was measured returning the generic Invalid request. Check your request parameters.. A max_tokens above the model's limit is documented as a generic 400. The errors table's 400 row now shows code: "400" with type: invalid_request_error. Tool.parameters gets additionalProperties: true. On deepseek-v4-flash a smaller max_tokens is raised to 16384, so a prompt within 16384 tokens of the window overflows (spec + deepseek card EN/ES).

Deploy note: the tool-parameters 400 is rolling out in a separate PR; merge this after that one is live.

4. deepseek-v4-flash window is 1048576 (deployment max_input_tokens set 2026-09-29): opencode limit.context, pi contextWindow and vscode maxInputTokens (1015808) updated from the one-token-stale 1048575, EN/ES; docsClientConfigs expects the new value.

5. Structured output guarantees + deepseek-v4-flash recipe. ResponseFormat and the deepseek-v4-flash card (EN/ES) say what each mode guarantees (json_object: valid JSON only, not the shape; json_schema strict: fields, types, required keys, on qwen3.6/gemma4). New examples.md section (EN/ES): one strict function tool forced with tool_choice, curl + Python; both snippets run as written on 2026-09-29 and returned arguments matching the schema. Tool.function declares strict.

Evidence (measured 2026-09-29, community synthetic key, read-only)

• POST /v1/responses deepseek-v4-flash, non-streaming: 200, status: completed, output = reasoning item + message item.
• Same with stream: true: 200, response.created at ~3.0s, reasoning summary deltas, response.output_text.delta spread from 3.3s to 5.5s, response.completed at the end.
• qwen3.6 / gemma4 with stream: true: 200, but all deltas and response.completed arrive within ~10ms at the end (32.5s / 56.9s).
• qwen3.8-flash: community proxy LiteLLM_ProxyModelTable max_input_tokens = 1048576; rate-limit hook window 1_048_576.

Tests

• openapiSpec.test.ts: 3 new describe blocks (responses models/streaming + Codex guide; qwen3.8-flash 1M across every surface in both locales; tool-parameters and context-overflow 400s). Mutation-checked: reverting each touched file to origin/main turns them red.
• docsClientConfigs.test.ts: expected qwen3.8-flash window 1_048_576 (nan#53 resolved).
• npm test: 66 files, 1320 tests passing. npm run build: complete. npx astro check: 0 errors.

🤖 Generated with Claude Code